Datasets:
MathVision seed variance: Qwen3.5-4B and Qwen3.5-9B
300 MathVision test problems × 5 sampling seeds × 2 models, thinking mode, with the full reasoning trace, exact prompt and generated token ids, and per-token log-probs (chosen + top-5) for every generation, so attention can be recomputed by teacher forcing.
Setup
| item | value |
|---|---|
| data | MathLLMs/MathVision@2837ddb test split (3,040); 300 sampled with random.Random(20261106) (156 multiple choice) → data/items.jsonl, data/images/ |
| models | Qwen/Qwen3.5-4B@851bf6e, Qwen/Qwen3.5-9B@c202236, bf16, 1×H200 each |
| serving | vLLM 0.31.1rc1.dev162+g4611c2c3d, default image preprocessing |
| decoding | thinking on; temperature 1.0, top-p 0.95, top-k 20, min-p 0, presence penalty 1.5, repetition 1.0 (model cards); max 38,912 new tokens |
| seeds | 84, 2026, 7, 11, 13 (per-request vLLM seed) |
| prompt | system prompt from the dataset's eval.yaml ("…on the final line output ONLY your answer in the form: ANSWER: "); user turn = image, then the question; multiple-choice options appended as Choices: <opt1>, <opt2>, … |
| scoring | last ANSWER: (fallback \boxed{}); free-form: numeric equality (fractions, \frac, degrees, units stripped) else normalized string; multiple choice: letter, or the reply's text mapped to exactly one option |
Scorer version. The prompt listed choices without letter labels (an oversight), so the
model often answered with the option text while the gold is a letter. outputs.jsonl
holds the original scoring (correct, v1 — undercounts multiple choice);
outputs_rescored.jsonl is the corrected scoring (correct, scorer: mathvision-v2,
predicted_letter; the old value is kept as correct_v1). 1,547 of 1,559 multiple-choice
replies map to a letter; the rest have no answer.
Results (scorer v2)
| Qwen3.5-4B | Qwen3.5-9B | |
|---|---|---|
| accuracy per seed (84 / 2026 / 7 / 11 / 13) | 70.9 / 71.6 / 71.2 / 71.6 / 71.9 | 76.3 / 71.3 / 72.3 / 73.7 / 74.0 |
| mean ± sd over seeds | 71.4 ± 0.3 | 73.5 ± 1.7 |
| items always correct / never / unstable | 56.9 / 13.4 / 29.8 % | 59.0 / 12.7 / 28.3 % |
| pass@5 / majority vote | 86.6 / 73.2 % | 87.3 / 77.3 % |
| mean pairwise correctness flip between seeds | 14.4 % | 13.4 % |
| mean completion tokens (correct / wrong) | 12,492 / 22,088 | 11,247 / 21,635 |
| length-capped / no answer (of 1,500) | 5 / 3 | 14 / 14 |
Aggregate accuracy is stable across seeds (sd ≤ 1.7 points) while ~29% of problems
change correctness between seeds. Wrong answers are ~1.8× longer. Both models are
always right on 150 and always wrong on 20 of the 299 common problems.
Per level, subject and type: analysis/variance_report_v2.json.
4B seed 13 has 299/300 results (mathvision_1595: one connection error), so cross-seed
statistics for 4B use the 299 common items.
Files
data/items.jsonl, data/images/ sampled problems (id, question, options, answer, level, subject, image sha256)
runs/main/<model>/seed-<s>/outputs.jsonl per request: messages, request params, reasoning, answer, prediction,
correct (v1), finish_reason, usage, latency, prompt_token_ids, output_token_ids
runs/main/<model>/seed-<s>/outputs_rescored.jsonl same rows, scorer v2 (use this)
runs/main/<model>/seed-<s>/token_logprobs.jsonl.gz per request: tokens, chosen-token logprob, top-5 [token, logprob] at every position
runs/main/<model>/seed-<s>/summary.json, run_manifests.jsonl
runs/main/<model>/attempts/<id>/ environment (pip freeze, nvidia-smi), vLLM server args/log, GPU telemetry, runner logs
runs/smoke/ 6-problem × 2-seed smoke runs
analysis/variance_report_v2.json all statistics above, per-item correct counts
code/ scoring, sampling, runner, analysis scripts used here
output_token_ids and the log-prob arrays cover every generated token (thinking +
answer). To get attention maps, feed prompt_token_ids + output_token_ids with the same
image (image_sha256) to the HF model with output_attentions=True.
- Downloads last month
- 595